What Do Deepfake Speech Detectors Actually Hear?
Abstract
Deepfake speech detectors often output a single score without explaining why an audio sample is flagged, where in the signal the evidence lies, or what cues drive the decision. We propose an audio-native explainability pipeline using Integrated Gradients on time-aligned self-supervised representations to localize decision evidence over time. We apply the proposed method to three WavLM-based detectors (AASIST, CA-MHFA, SLS) on ASVspoof 5 and manually annotate the highest-attribution regions to provide a semantic meaning of the most important cues. Despite similar performance, the detectors rely on different cues: AASIST emphasizes non-speech/environment cues, CA-MHFA focuses on localized phoneme artifacts, and SLS relies on word boundaries and spectral integrity. We move beyond speculative reasoning and validate our findings by causal masking of the primary detector cues. Observed performance degradation further supports the explained detector semantics.
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
The paper presents a novel approach to explainability in deepfake speech detection, offering valuable insights into the decision-making processes of various detectors. The comprehensive methodology and rigorous experimental evaluation contribute significantly to the field, paving the way for future advancements in audio forensics and detector design.
Methodology
The paper introduces an innovative audio-native explainability pipeline using Integrated Gradients (IG) to analyze deepfake speech detectors. The methodology is robust, leveraging self-supervised learning (SSL) representations and providing a structured annotation protocol for semantic analysis. The use of causal masking to validate findings adds rigor to the approach, allowing for a clear understanding of the decision-making processes of different detectors.
Results & evaluation
The experiments are well-structured, utilizing the ASVspoof 5 dataset and a representative subset of recordings to evaluate the detectors. The results reveal distinct cue reliance among the detectors, supported by qualitative annotations and quantitative performance metrics. The findings highlight the complementary nature of the detectors, suggesting potential for ensemble methods.
Reproducibility
The paper provides sufficient implementation details, including the use of the Captum library for IG calculations and a clear description of the annotation protocol. The availability of a GitHub repository enhances reproducibility, allowing other researchers to replicate the study and build upon the findings.
Limitations
One limitation is the focus on SSL-based models, which may restrict the generalizability of the findings to other architectures. Additionally, the reliance on human annotations, while valuable, introduces subjectivity that could affect the consistency of the results. The shared vulnerability to audio compression across detectors also suggests a potential weakness in their robustness.
Broader impact
The findings have significant implications for audio forensics and the development of deepfake detection systems. By providing insights into the decision-making processes of detectors, the work can inform the design of more effective systems and enhance the interpretability of AI in critical applications. The potential for ensemble methods could lead to advancements in the field, improving detection accuracy and reliability. The paper presents a novel approach to explainability in deepfake speech detection, offering valuable insights into the decision-making processes of various detectors. The comprehensive methodology and rigorous experimental evaluation contribute significantly to the field, paving the way for future advancements in audio forensics and detector design.
Author of this paper? Send a correction.